|
1
|
1 |
|
from keras.models import Sequential |
|
2
|
1 |
|
from keras.layers import Dense, Activation, Convolution1D, Lambda, Convolution2D, Flatten, \ |
|
|
|
|
|
|
3
|
|
|
Reshape, LSTM, Dropout, TimeDistributed, BatchNormalization |
|
4
|
1 |
|
from keras.regularizers import l2 |
|
5
|
1 |
|
from keras.optimizers import Adam |
|
6
|
1 |
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
1 |
|
def generate_models( |
|
10
|
|
|
x_shape, number_of_classes, number_of_models=5, model_type=None, |
|
11
|
|
|
cnn_min_layers=1, cnn_max_layers=10, |
|
12
|
|
|
cnn_min_filters=10, cnn_max_filters=100, |
|
13
|
|
|
cnn_min_fc_nodes=10, cnn_max_fc_nodes=2000, |
|
14
|
|
|
deepconvlstm_min_conv_layers=1, deepconvlstm_max_conv_layers=10, |
|
15
|
|
|
deepconvlstm_min_conv_filters=10, deepconvlstm_max_conv_filters=100, |
|
16
|
|
|
deepconvlstm_min_lstm_layers=1, deepconvlstm_max_lstm_layers=5, |
|
17
|
|
|
deepconvlstm_min_lstm_dims=10, deepconvlstm_max_lstm_dims=100, |
|
18
|
|
|
low_lr=1, high_lr=4, low_reg=1, high_reg=4 |
|
19
|
|
|
): |
|
20
|
|
|
""" |
|
21
|
|
|
Generate one or multiple Keras models with random hyperparameters. |
|
22
|
|
|
|
|
23
|
|
|
Parameters |
|
24
|
|
|
---------- |
|
25
|
|
|
x_shape |
|
26
|
|
|
Shape of the input dataset: (num_samples, num_timesteps, num_channels) |
|
27
|
|
|
number_of_classes |
|
28
|
|
|
Number of classes for classification task |
|
29
|
|
|
number_of_models |
|
30
|
|
|
Number of models to generate |
|
31
|
|
|
model_type : str (optional) |
|
32
|
|
|
Type of model to build: 'CNN' or 'DeepConvLSTM'. Default option None generates both models. |
|
|
|
|
|
|
33
|
|
|
cnn_min_layers : int |
|
34
|
|
|
minimum of Conv layers in CNN model |
|
35
|
|
|
cnn_max_layers : int |
|
36
|
|
|
maximum of Conv layers in CNN model |
|
37
|
|
|
cnn_min_filters : int |
|
38
|
|
|
minimum number of filters per Conv layer in CNN model |
|
39
|
|
|
cnn_max_filters : int |
|
40
|
|
|
maximum number of filters per Conv layer in CNN model |
|
41
|
|
|
cnn_min_fc_nodes : int |
|
42
|
|
|
minimum number of hidden nodes per Dense layer in CNN model |
|
43
|
|
|
cnn_max_fc_nodes : int |
|
44
|
|
|
maximum number of hidden nodes per Dense layer in CNN model |
|
45
|
|
|
deepconvlstm_min_conv_layers : int |
|
46
|
|
|
minimum number of Conv layers in DeepConvLSTM model |
|
47
|
|
|
deepconvlstm_max_conv_layers : int |
|
48
|
|
|
maximum number of Conv layers in DeepConvLSTM model |
|
49
|
|
|
deepconvlstm_min_conv_filters : int |
|
50
|
|
|
minimum number of filters per Conv layer in DeepConvLSTM model |
|
51
|
|
|
deepconvlstm_max_conv_filters : int |
|
52
|
|
|
maximum number of filters per Conv layer in DeepConvLSTM model |
|
53
|
|
|
deepconvlstm_min_lstm_layers : int |
|
54
|
|
|
minimum number of Conv layers in DeepConvLSTM model |
|
55
|
|
|
deepconvlstm_max_lstm_layers : int |
|
56
|
|
|
maximum number of Conv layers in DeepConvLSTM model |
|
57
|
|
|
deepconvlstm_min_lstm_dims : int |
|
58
|
|
|
minimum number of hidden nodes per LSTM layer in DeepConvLSTM model |
|
59
|
|
|
deepconvlstm_max_lstm_dims : int |
|
60
|
|
|
maximum number of hidden nodes per LSTM layer in DeepConvLSTM model |
|
61
|
|
|
low_lr : float |
|
62
|
|
|
minimum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` and `10**(-high_reg)` |
|
|
|
|
|
|
63
|
|
|
high_lr : float |
|
64
|
|
|
maximum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
65
|
|
|
and `10**(-high_reg)` |
|
66
|
|
|
low_reg : float |
|
67
|
|
|
minimum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
68
|
|
|
and `10**(-high_reg)` |
|
69
|
|
|
high_reg : float |
|
70
|
|
|
maximum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
71
|
|
|
and `10**(-high_reg)` |
|
72
|
|
|
|
|
73
|
|
|
Returns |
|
74
|
|
|
------- |
|
75
|
|
|
List of compiled models |
|
76
|
|
|
""" |
|
77
|
1 |
|
models = [] |
|
78
|
1 |
|
for _ in range(0, number_of_models): |
|
79
|
1 |
|
if model_type is None: # random model choice: |
|
80
|
1 |
|
current_model_type = 'CNN' if np.random.random( |
|
81
|
|
|
) < 0.5 else 'DeepConvLSTM' |
|
82
|
|
|
else: # user-defined model choice: |
|
83
|
|
|
current_model_type = model_type |
|
84
|
1 |
|
generate_model = None |
|
85
|
1 |
|
if current_model_type == 'CNN': |
|
86
|
1 |
|
generate_model = generate_CNN_model # object is a function |
|
87
|
1 |
|
hyperparameters = generate_CNN_hyperparameter_set( |
|
88
|
|
|
min_layers=cnn_min_layers, max_layers=cnn_max_layers, |
|
89
|
|
|
min_filters=cnn_min_filters, max_filters=cnn_max_filters, |
|
90
|
|
|
min_fc_nodes=cnn_min_fc_nodes, max_fc_nodes=cnn_max_fc_nodes, |
|
91
|
|
|
low_lr=low_lr, high_lr=high_lr, low_reg=low_reg, high_reg=high_reg) |
|
|
|
|
|
|
92
|
1 |
|
if current_model_type == 'DeepConvLSTM': |
|
93
|
|
|
generate_model = generate_DeepConvLSTM_model # object is a function |
|
|
|
|
|
|
94
|
|
|
hyperparameters = generate_DeepConvLSTM_hyperparameter_set( |
|
95
|
|
|
min_conv_layers=deepconvlstm_min_conv_layers, max_conv_layers=deepconvlstm_max_conv_layers, |
|
|
|
|
|
|
96
|
|
|
min_conv_filters=deepconvlstm_min_conv_filters, max_conv_filters=deepconvlstm_max_conv_filters, |
|
|
|
|
|
|
97
|
|
|
min_lstm_layers=deepconvlstm_min_lstm_layers, max_lstm_layers=deepconvlstm_max_lstm_layers, |
|
|
|
|
|
|
98
|
|
|
min_lstm_dims=deepconvlstm_min_lstm_dims, max_lstm_dims=deepconvlstm_max_lstm_dims, |
|
|
|
|
|
|
99
|
|
|
low_lr=low_lr, high_lr=high_lr, low_reg=low_reg, high_reg=high_reg) |
|
|
|
|
|
|
100
|
1 |
|
models.append( |
|
101
|
|
|
(generate_model(x_shape, number_of_classes, **hyperparameters), hyperparameters, current_model_type)) |
|
|
|
|
|
|
102
|
1 |
|
return models |
|
103
|
|
|
|
|
104
|
|
|
|
|
105
|
1 |
|
def generate_DeepConvLSTM_model( |
|
|
|
|
|
|
106
|
|
|
x_shape, class_number, filters, lstm_dims, learning_rate=0.01, |
|
107
|
|
|
regularization_rate=0.01): |
|
108
|
|
|
""" |
|
109
|
|
|
Generate a model with convolution and LSTM layers. |
|
110
|
|
|
See Ordonez et al., 2016, http://dx.doi.org/10.3390/s16010115 |
|
111
|
|
|
|
|
112
|
|
|
Parameters |
|
113
|
|
|
---------- |
|
114
|
|
|
x_shape : tuple |
|
115
|
|
|
Shape of the input dataset: (num_samples, num_timesteps, num_channels) |
|
116
|
|
|
class_number : int |
|
117
|
|
|
Number of classes for classification task |
|
118
|
|
|
filters : list of ints |
|
119
|
|
|
number of filters for each convolutional layer |
|
120
|
|
|
lstm_dims : list of ints |
|
121
|
|
|
number of hidden nodes for each LSTM layer |
|
122
|
|
|
learning_rate : float |
|
123
|
|
|
regularization_rate : float |
|
124
|
|
|
|
|
125
|
|
|
Returns |
|
126
|
|
|
------- |
|
127
|
|
|
The compiled Keras model |
|
128
|
|
|
""" |
|
129
|
1 |
|
dim_length = x_shape[1] # number of samples in a time series |
|
130
|
1 |
|
dim_channels = x_shape[2] # number of channels |
|
131
|
1 |
|
output_dim = class_number # number of classes |
|
132
|
1 |
|
weightinit = 'lecun_uniform' # weight initialization |
|
133
|
1 |
|
model = Sequential() # initialize model |
|
134
|
1 |
|
model.add(BatchNormalization(input_shape=(dim_length, dim_channels))) |
|
135
|
|
|
# reshape a 2 dimensional array per file/person/object into a |
|
136
|
|
|
# 3 dimensional array |
|
137
|
1 |
|
model.add( |
|
138
|
|
|
Reshape(target_shape=(1, dim_length, dim_channels))) |
|
139
|
1 |
|
for filt in filters: |
|
140
|
|
|
# filt: number of filters used in a layer |
|
141
|
|
|
# filters: vector of filt values |
|
142
|
1 |
|
model.add( |
|
143
|
|
|
Convolution2D(filt, nb_row=3, nb_col=1, border_mode='same', |
|
144
|
|
|
W_regularizer=l2(regularization_rate), init=weightinit)) |
|
|
|
|
|
|
145
|
1 |
|
model.add(BatchNormalization()) |
|
146
|
1 |
|
model.add(Activation('relu')) |
|
147
|
|
|
# reshape 3 dimensional array back into a 2 dimensional array, |
|
148
|
|
|
# but now with more dept as we have the the filters for each channel |
|
149
|
1 |
|
model.add(Reshape(target_shape=(dim_length, filters[-1] * dim_channels))) |
|
150
|
|
|
|
|
151
|
1 |
|
for lstm_dim in lstm_dims: |
|
152
|
1 |
|
model.add(LSTM(output_dim=lstm_dim, return_sequences=True, |
|
153
|
|
|
activation='tanh')) |
|
154
|
|
|
|
|
155
|
1 |
|
model.add(Dropout(0.5)) # dropout before the dense layer |
|
156
|
|
|
# set up final dense layer such that every timestamp is given one |
|
157
|
|
|
# classification |
|
158
|
1 |
|
model.add( |
|
159
|
|
|
TimeDistributed(Dense(output_dim, W_regularizer=l2(regularization_rate)))) |
|
|
|
|
|
|
160
|
1 |
|
model.add(Activation("softmax")) |
|
161
|
|
|
# Final classification layer - per timestep |
|
162
|
1 |
|
model.add(Lambda(lambda x: x[:, -1, :], output_shape=[output_dim])) |
|
163
|
|
|
|
|
164
|
1 |
|
model.compile(loss='categorical_crossentropy', |
|
165
|
|
|
optimizer=Adam(lr=learning_rate), |
|
166
|
|
|
metrics=['accuracy']) |
|
167
|
|
|
|
|
168
|
1 |
|
return model |
|
169
|
|
|
|
|
170
|
|
|
|
|
171
|
1 |
|
def generate_CNN_model(x_shape, class_number, filters, fc_hidden_nodes, |
|
|
|
|
|
|
172
|
|
|
learning_rate=0.01, regularization_rate=0.01): |
|
173
|
|
|
""" |
|
174
|
|
|
Generate a convolutional neural network (CNN) model. |
|
175
|
|
|
|
|
176
|
|
|
The compiled Keras model is returned. |
|
177
|
|
|
|
|
178
|
|
|
Parameters |
|
179
|
|
|
---------- |
|
180
|
|
|
x_shape : tuple |
|
181
|
|
|
Shape of the input dataset: (num_samples, num_timesteps, num_channels) |
|
182
|
|
|
class_number : int |
|
183
|
|
|
Number of classes for classification task |
|
184
|
|
|
filters : list of ints |
|
185
|
|
|
number of filters for each convolutional layer |
|
186
|
|
|
fc_hidden_nodes : int |
|
187
|
|
|
number of hidden nodes for the hidden dense layer |
|
188
|
|
|
learning_rate : float |
|
189
|
|
|
regularization_rate : float |
|
190
|
|
|
|
|
191
|
|
|
Returns |
|
192
|
|
|
------- |
|
193
|
|
|
The compiled Keras model |
|
194
|
|
|
""" |
|
195
|
1 |
|
dim_length = x_shape[1] # number of samples in a time series |
|
196
|
1 |
|
dim_channels = x_shape[2] # number of channels |
|
197
|
1 |
|
outputdim = class_number # number of classes |
|
198
|
1 |
|
weightinit = 'lecun_uniform' # weight initialization |
|
199
|
1 |
|
model = Sequential() |
|
200
|
1 |
|
model.add( |
|
201
|
|
|
BatchNormalization( |
|
202
|
|
|
input_shape=( |
|
203
|
|
|
dim_length, |
|
204
|
|
|
dim_channels), |
|
205
|
|
|
mode=0, |
|
206
|
|
|
axis=2)) |
|
207
|
1 |
|
for filter_number in filters: |
|
208
|
1 |
|
model.add(Convolution1D(filter_number, 3, border_mode='same', |
|
209
|
|
|
W_regularizer=l2(regularization_rate), init=weightinit)) |
|
|
|
|
|
|
210
|
1 |
|
model.add(BatchNormalization()) |
|
211
|
1 |
|
model.add(Activation('relu')) |
|
212
|
1 |
|
model.add(Flatten()) |
|
213
|
1 |
|
model.add(Dense(output_dim=fc_hidden_nodes, |
|
214
|
|
|
W_regularizer=l2(regularization_rate), init=weightinit)) # Fully connected layer |
|
|
|
|
|
|
215
|
1 |
|
model.add(Activation('relu')) # Relu activation |
|
216
|
1 |
|
model.add(Dense(output_dim=outputdim, init=weightinit)) |
|
217
|
1 |
|
model.add(BatchNormalization()) |
|
218
|
1 |
|
model.add(Activation("softmax")) # Final classification layer |
|
219
|
|
|
|
|
220
|
1 |
|
model.compile(loss='categorical_crossentropy', |
|
221
|
|
|
optimizer=Adam(lr=learning_rate), |
|
222
|
|
|
metrics=['accuracy']) |
|
223
|
|
|
|
|
224
|
1 |
|
return model |
|
225
|
|
|
|
|
226
|
|
|
|
|
227
|
1 |
|
def generate_CNN_hyperparameter_set(min_layers=1, max_layers=10, |
|
|
|
|
|
|
228
|
|
|
min_filters=10, max_filters=100, |
|
229
|
|
|
min_fc_nodes=10, max_fc_nodes=2000, |
|
230
|
|
|
low_lr=1, high_lr=4, low_reg=1, high_reg=4): |
|
|
|
|
|
|
231
|
|
|
""" Generate a hyperparameter set that define a CNN model. |
|
232
|
|
|
|
|
233
|
|
|
Parameters |
|
234
|
|
|
---------- |
|
235
|
|
|
min_layers : int |
|
236
|
|
|
minimum of Conv layers |
|
237
|
|
|
max_layers : int |
|
238
|
|
|
maximum of Conv layers |
|
239
|
|
|
min_filters : int |
|
240
|
|
|
minimum number of filters per Conv layer |
|
241
|
|
|
max_filters : int |
|
242
|
|
|
maximum number of filters per Conv layer |
|
243
|
|
|
min_fc_nodes : int |
|
244
|
|
|
minimum number of hidden nodes per Dense layer |
|
245
|
|
|
max_fc_nodes : int |
|
246
|
|
|
maximum number of hidden nodes per Dense layer |
|
247
|
|
|
low_lr : float |
|
248
|
|
|
minimum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
249
|
|
|
and `10**(-high_reg)` |
|
250
|
|
|
high_lr : float |
|
251
|
|
|
maximum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
252
|
|
|
and `10**(-high_reg)` |
|
253
|
|
|
low_reg : float |
|
254
|
|
|
minimum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
255
|
|
|
and `10**(-high_reg)` |
|
256
|
|
|
high_reg : float |
|
257
|
|
|
maximum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
258
|
|
|
and `10**(-high_reg)` |
|
259
|
|
|
|
|
260
|
|
|
Returns |
|
261
|
|
|
---------- |
|
262
|
|
|
hyperparameters : dict |
|
263
|
|
|
parameters for a CNN model |
|
264
|
|
|
""" |
|
265
|
1 |
|
hyperparameters = generate_base_hyper_parameter_set( |
|
266
|
|
|
low_lr, high_lr, low_reg, high_reg) |
|
267
|
1 |
|
number_of_layers = np.random.randint(min_layers, max_layers + 1) |
|
268
|
1 |
|
hyperparameters['filters'] = np.random.randint( |
|
269
|
|
|
min_filters, max_filters + 1, number_of_layers) |
|
270
|
1 |
|
hyperparameters['fc_hidden_nodes'] = np.random.randint( |
|
271
|
|
|
min_fc_nodes, max_fc_nodes + 1) |
|
272
|
1 |
|
return hyperparameters |
|
273
|
|
|
|
|
274
|
|
|
|
|
275
|
1 |
|
def generate_DeepConvLSTM_hyperparameter_set( |
|
|
|
|
|
|
276
|
|
|
min_conv_layers=1, max_conv_layers=10, |
|
277
|
|
|
min_conv_filters=10, max_conv_filters=100, |
|
278
|
|
|
min_lstm_layers=1, max_lstm_layers=5, |
|
279
|
|
|
min_lstm_dims=10, max_lstm_dims=100, |
|
280
|
|
|
low_lr=1, high_lr=4, low_reg=1, high_reg=4): |
|
281
|
|
|
""" Generate a hyperparameter set that defines a DeepConvLSTM model. |
|
282
|
|
|
|
|
283
|
|
|
Parameters |
|
284
|
|
|
---------- |
|
285
|
|
|
min_conv_layers : int |
|
286
|
|
|
minimum number of Conv layers in DeepConvLSTM model |
|
287
|
|
|
max_conv_layers : int |
|
288
|
|
|
maximum number of Conv layers in DeepConvLSTM model |
|
289
|
|
|
min_conv_filters : int |
|
290
|
|
|
minimum number of filters per Conv layer in DeepConvLSTM model |
|
291
|
|
|
max_conv_filters : int |
|
292
|
|
|
maximum number of filters per Conv layer in DeepConvLSTM model |
|
293
|
|
|
min_lstm_layers : int |
|
294
|
|
|
minimum number of Conv layers in DeepConvLSTM model |
|
295
|
|
|
max_lstm_layers : int |
|
296
|
|
|
maximum number of Conv layers in DeepConvLSTM model |
|
297
|
|
|
min_lstm_dims : int |
|
298
|
|
|
minimum number of hidden nodes per LSTM layer in DeepConvLSTM model |
|
299
|
|
|
max_lstm_dims : int |
|
300
|
|
|
maximum number of hidden nodes per LSTM layer in DeepConvLSTM model |
|
301
|
|
|
low_lr : float |
|
302
|
|
|
minimum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
303
|
|
|
and `10**(-high_reg)` |
|
304
|
|
|
high_lr : float |
|
305
|
|
|
maximum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
306
|
|
|
and `10**(-high_reg)` |
|
307
|
|
|
low_reg : float |
|
308
|
|
|
minimum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
309
|
|
|
and `10**(-high_reg)` |
|
310
|
|
|
high_reg : float |
|
311
|
|
|
maximum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
312
|
|
|
and `10**(-high_reg)` |
|
313
|
|
|
|
|
314
|
|
|
Returns |
|
315
|
|
|
---------- |
|
316
|
|
|
hyperparameters: dict |
|
317
|
|
|
hyperparameters for a DeepConvLSTM model |
|
318
|
|
|
""" |
|
319
|
1 |
|
hyperparameters = generate_base_hyper_parameter_set( |
|
320
|
|
|
low_lr, high_lr, low_reg, high_reg) |
|
321
|
1 |
|
number_of_conv_layers = np.random.randint( |
|
322
|
|
|
min_conv_layers, max_conv_layers + 1) |
|
323
|
1 |
|
hyperparameters['filters'] = np.random.randint( |
|
324
|
|
|
min_conv_filters, max_conv_filters + 1, number_of_conv_layers) |
|
325
|
1 |
|
number_of_lstm_layers = np.random.randint( |
|
326
|
|
|
min_lstm_layers, max_lstm_layers + 1) |
|
327
|
1 |
|
hyperparameters['lstm_dims'] = np.random.randint( |
|
328
|
|
|
min_lstm_dims, max_lstm_dims + 1, number_of_lstm_layers) |
|
329
|
1 |
|
return hyperparameters |
|
330
|
|
|
|
|
331
|
|
|
|
|
332
|
1 |
|
def generate_base_hyper_parameter_set( |
|
|
|
|
|
|
333
|
|
|
low_lr=1, |
|
334
|
|
|
high_lr=4, |
|
335
|
|
|
low_reg=1, |
|
336
|
|
|
high_reg=4): |
|
337
|
|
|
""" Generate a base set of hyperparameters that are necessary for any model, but sufficient for none. |
|
|
|
|
|
|
338
|
|
|
|
|
339
|
|
|
Parameters |
|
340
|
|
|
---------- |
|
341
|
|
|
low_lr : float |
|
342
|
|
|
minimum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` and `10**(-high_reg)` |
|
|
|
|
|
|
343
|
|
|
high_lr : float |
|
344
|
|
|
maximum of log range for learning rate: learning rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
345
|
|
|
and `10**(-high_reg)` |
|
346
|
|
|
low_reg : float |
|
347
|
|
|
minimum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
348
|
|
|
and `10**(-high_reg)` |
|
349
|
|
|
high_reg : float |
|
350
|
|
|
maximum of log range for regularization rate: regularization rate is sampled between `10**(-low_reg)` |
|
|
|
|
|
|
351
|
|
|
and `10**(-high_reg)` |
|
352
|
|
|
|
|
353
|
|
|
Returns |
|
354
|
|
|
------- |
|
355
|
|
|
hyperparameters : dict |
|
356
|
|
|
basis hyperpameters |
|
357
|
|
|
""" |
|
358
|
1 |
|
hyperparameters = {} |
|
359
|
1 |
|
hyperparameters['learning_rate'] = get_learning_rate(low_lr, high_lr) |
|
360
|
1 |
|
hyperparameters['regularization_rate'] = get_regularization( |
|
361
|
|
|
low_reg, high_reg) |
|
362
|
1 |
|
return hyperparameters |
|
363
|
|
|
|
|
364
|
|
|
|
|
365
|
1 |
|
def get_learning_rate(low=1, high=4): |
|
366
|
|
|
""" Return random learning rate 10^-n where n is sampled uniformly between low and high bounds. |
|
|
|
|
|
|
367
|
|
|
|
|
368
|
|
|
Parameters |
|
369
|
|
|
---------- |
|
370
|
|
|
low : float |
|
371
|
|
|
low bound |
|
372
|
|
|
high : float |
|
373
|
|
|
high bound |
|
374
|
|
|
""" |
|
375
|
1 |
|
result = 10 ** (-np.random.uniform(low, high)) |
|
376
|
1 |
|
return result |
|
377
|
|
|
|
|
378
|
|
|
|
|
379
|
1 |
|
def get_regularization(low=1, high=4): |
|
380
|
|
|
""" Return random regularization rate 10^-n where n is sampled uniformly between low and high bounds. |
|
|
|
|
|
|
381
|
|
|
|
|
382
|
|
|
Parameters |
|
383
|
|
|
---------- |
|
384
|
|
|
low : float |
|
385
|
|
|
low bound |
|
386
|
|
|
high : float |
|
387
|
|
|
high bound |
|
388
|
|
|
""" |
|
389
|
|
|
return 10 ** (-np.random.uniform(low, high)) |
|
390
|
|
|
|
NLeSC /
mcfly
Loading history...
This check looks for lines that are too long. You can specify the maximum line length.